Public research scaffold. Release status: scaffolded. License posture requires human review.
Energy Per Intelligence (EPI) is Francisco Abner Rivera's public research surface for a simple engineering question:
How much energy does one unit of useful model output cost?
This repository frames EPI as a measurement and reporting discipline for AI systems. It is part of the Franzabner public technical brand, not a released benchmark, not a dataset release, not a model release, and not a Hugging Face artifact.
| Item | Status |
|---|---|
| Public status | Research scaffold |
| Release status | Scaffolded |
| Measurement data | Not released |
| Benchmark results | Not released |
| Model weights | Not released |
| Hugging Face model, dataset, or Space | Not created by this repo |
| License posture | Existing license files are unchanged; human review required before any license change or external reliance |
EPI is a proposed metric for comparing energy cost against output usefulness:
EPI = joules per output unit / reviewed quality score
The exact denominator must be selected by a reviewed evaluation protocol. That can be a benchmark score, task score, or domain-specific rubric, but this repository does not claim validated benchmark results or measured production values.
Lower EPI means less energy per useful output unit under the selected protocol. The public work here is the method, status discipline, and research framing needed before measured claims are made.
This repo connects to the live Franzabner public proof stack:
| Repo | Role |
|---|---|
| franzabner-proof-stack | Public navigation and proof-routing spine |
| hf-card-templates | Hugging Face release-readiness templates and boundary gates |
| epi-bench | Planned EPI calculation and report tooling scaffold |
| epi-meter | Planned public-safe AC-side measurement instrument scaffold |
- EPI metric framing.
- Public-safe research structure.
- Placeholder data directories.
- Skeleton calculation and visualization code.
- Documentation for future measurement review.
- Boundary language for public, private, and sealed material.
This repository does not claim:
- released benchmark results;
- validated EPI scores;
- released datasets;
- hosted models;
- hosted Hugging Face datasets, models, or Spaces;
- deployed systems;
- client or customer use;
- revenue outcomes;
- production readiness;
- private YOSO-YAi model, corpus, endpoint, or infrastructure disclosure.
Human review is required before:
- publishing measured EPI results;
- publishing raw traces or benchmark outputs;
- linking to external Hugging Face artifacts;
- changing license posture;
- claiming release, deployment, client usage, or benchmark validity.
Public examples must be synthetic or explicitly approved for publication. Private corpora, private model weights, private infrastructure, private endpoints, customer files, and sealed company implementation stay out of this repository.
Measure the energy. Preserve the boundary. Do not make a claim before the evidence is public.